A semi-supervised panoramic X-ray target detection method based on synthetic data

Abstract

This invention discloses a semi-supervised panoramic X-ray target detection method based on synthetic data. First, a panoramic X-ray dataset is acquired, and a controllable synthetic model is trained to generate class-balanced samples. Second, a panoramic X-ray dataset and diagnostic annotations are acquired, and a supervised target detection and diagnosis model is trained. Then, weak labels from the synthetic samples are acquired, and a semi-supervised target detection method is used to train a diagnostic model to achieve more accurate diagnostic performance. Finally, the trained target detection and diagnosis model is deployed in the software backend of a terminal device, and the patient's panoramic X-ray data is imported to make a diagnosis. This invention solves the problem of insufficient model accuracy caused by insufficient data and insufficient data annotation in traditional methods.

Description

Technical Field

[0001] This invention belongs to the field of image target detection technology in computer vision, and relates to a target detection method for medical images (panoramic X-rays), particularly to a semi-supervised panoramic X-ray target detection method and system based on synthetic data, which can be specifically applied to the field of auxiliary diagnosis of periodontitis. Background Technology

[0002] Periodontitis is one of the most common oral diseases, affecting over 80% of adults in some form. Mild periodontitis can lead to bleeding or loss of tissue attachment, while more severe cases can cause irreversible bone damage or even tooth loss. Analysis of panoramic X-rays is a primary basis for periodontitis staging, with the degree of alveolar bone loss being a key diagnostic indicator, measured by interpreting the grayscale information of the teeth on the panoramic X-ray. While panoramic X-ray analysis is crucial, interpreting the images is time-consuming and laborious for doctors, necessitating the use of artificial intelligence to assist or even replace doctors in periodontitis staging diagnoses. However, current challenges stem primarily from two aspects: firstly, insufficient image data—most hospitals lack sufficient data to support the training of deep learning models; and secondly, insufficient data annotation—doctors spend even more time and effort annotating data than interpreting the images, and the limited annotation available is insufficient for training deep learning models. This invention addresses these two engineering challenges by providing solutions that enhance the accuracy of panoramic X-ray target detection methods, demonstrating broad development prospects. Summary of the Invention

[0003] This invention belongs to the field of image target detection technology in computer vision, and relates to a target detection method and system for medical images (panoramic X-rays), particularly to a semi-supervised panoramic X-ray target detection method and system based on synthetic data. Specifically, it can be applied to the field of auxiliary diagnosis of periodontitis, and solves the problem of insufficient model accuracy caused by insufficient image data and insufficient data annotation in traditional methods.

[0004] The technical solution adopted by the method of the present invention is: a semi-supervised panoramic X-ray target detection method based on synthetic data, comprising the following steps:

[0005] Step 1: Obtain the panoramic X-ray dataset, train the controllable synthesis model, and control the depth synthesis model to generate class-balanced samples; including the following sub-steps:

[0006] Step 1.1: Prepare a small number of labeled panoramic X-ray images. Encode each target in each sample according to the disease label category. After weighting, obtain the overall disease label as the category label of the entire sample.

[0007] Step 1.2: Train a deep synthesis model using the dataset, requiring the synthesis results to satisfy the category while synthesizing high-quality results with high realism;

[0008] Step 1.3: Apply the trained deep synthetic model to synthesize synthetic data that is more than K times larger than the original labeled dataset, while also ensuring class balance in the synthetic data.

[0009] Step 2: Obtain the panoramic X-ray dataset and diagnostic annotations, and train the deep object detection network model;

[0010] Step 3: Obtain weak labels for the synthetic samples, and train the diagnostic model using a semi-supervised object detection method on the entire unlabeled dataset to achieve more accurate diagnostic performance; this includes the following sub-steps:

[0011] Step 3.1: Using the target detection and diagnosis model obtained in Step 2, apply it to the unlabeled samples synthesized in Step 1 to obtain weak labels;

[0012] Step 3.2: Train a semi-supervised object detection model by introducing noise into the input samples and using consistency regularization to train the semi-supervised object detection model to convergence on a dataset with weak labels, thus obtaining a preliminary semi-supervised detection model.

[0013] Step 3.3: Use the semi-supervised detection model obtained in step 3.2 as the target detection and diagnosis model in step 3.1, re-acquire weak labels, and repeat steps 3.1 and 3.2. This iterative process is repeated several times until the model fully converges, and the final semi-supervised detection model is obtained.

[0014] Step 4: Deploy the trained target detection and diagnosis model in the software backend of the terminal device and import the patient's panoramic X-ray data to make a diagnosis.

[0015] Furthermore, the specific implementation method of step 1.1 is as follows;

[0016] Category labels are defined as severity coefficients (SC):

[0017]

[0018] In the formula C i W represents the number of categories corresponding to the i-th target, and N is the number of targets in a single sample; i A coefficient assigned based on the number of categories.

[0019] Furthermore, the deep synthesis model is an improved generative adversarial network, including a generator and a discriminator. The specific processing procedure of the generative adversarial network is as follows: given an SC value, along with a style vector S and constant inputs, it is used as the generator input. These inputs are processed by the generator to output a synthetic sample, i.e., a fake sample with the SC category. The style vector S is obtained by mapping a random vector Z, and the constant input is arbitrary. The synthetic sample is input into the discriminator to obtain its authenticity and SC category discrimination values, respectively. The discrimination value and the given SC value are used to calculate a loss function to guide the training of the generator and the discriminator. Finally, a generator that can control the synthesis of a given SC value is obtained.

[0020] Furthermore, the discriminator has two branches. The input sample to be identified passes through the upper branch, traversing several residual downsampling blocks D to reduce the feature map size; and through the lower branch, traversing a simple nearest neighbor downsampling block P to further reduce the feature map size. After each downsampling stage, the two branches perform feature map addition, fusing the results as the input for the next stage. The final fused features then pass through another residual downsampling block before outputting the prediction result. The P block uses a 2x2 nearest neighbor downsampling operation. In the D block, the input first undergoes a 1x1 convolution, then splits into two paths: one branch performs two 3x3 convolutions, and the other performs a 1x1 convolution. The two paths then perform feature map addition, and finally, a 2x2 nearest neighbor downsampling operation is performed before outputting the result.

[0021] Furthermore, the generator comprises two parts: an upper module that maps the input vector Z to a style vector S through multiple fully connected layers, and a lower module that controls the input generation result using the style vector. In the main module, the low-resolution input is expanded through multiple upsampling feature extraction blocks B, the process of which is controlled by the components of the style vector S. Then, the output of each stage of block B is output as OUT1, ..., OUTn through the style mapping output block R. The sum of all these outputs is the final output result, where n represents the number of style vector components. In the structure of block B, the input feature map is upsampled, then subjected to two 3x3 convolutions, and finally output. In the structure of block R, the input feature map is subjected to a 3x3 convolution, then upsampled, and finally output.

[0022] Furthermore, the specific calculation process of the loss function is as follows;

[0023] The original definition of the loss function in GAN is incorporated into the severity coefficient control term, as shown below for the losses of the generator and discriminator, respectively:

[0024]

[0025]

[0026] Among them, Ez～p(z) [] represents expectation, P z (z) is the distribution of the values ​​of z, P data (x) is the distribution of x taking values ​​of real samples (data); z is the input random vector, and x is a real sample; the training objective of the discriminator D is to maximize the distribution distance between real samples and generated samples, while the training objective of the generator G is to minimize the distribution distance to mislead the discriminator;

[0027] An additional path length regularization term is introduced for training. This is done by interpolating different input vectors to obtain results that satisfy a linear distance relationship, defined as follows:

[0028] Reg spl =λ(||G(z')-y||2-α) 2

[0029] In the formula, z' is the sample of the input style vector; y is the image generated from z input; α is the average path constant of the sample set; and λ is the coefficient of the set regularization intensity.

[0030] In summary, the loss function of the improved generative adversarial network is defined as follows:

[0031] L = Loss D +Loss G +Reg spl .

[0032] Furthermore, the specific implementation of step 2 includes the following sub-steps:

[0033] Step 2.1: Prepare a small dataset of labeled panoramic X-ray images, and encode each label according to the sample to obtain the target that the depth target detection network model needs to fit.

[0034] Step 2.2: Based on the deviation between the fitted target and the target output by the depth target detection network model, train the network parameters to make the depth target detection network model fit the target.

[0035] Furthermore, the deep object detection network model is either a one-stage detection model or a two-stage detection model based on convolutional neural networks.

[0036] Furthermore, the network model structure of the semi-supervised target detection model is the same as that of the deep target detection network model in step 2.

[0037] Compared with existing technologies, the semi-supervised panoramic X-ray target detection method and system based on synthetic data proposed in this invention has the following advantages:

[0038] 1) It is easier to obtain data. Compared with traditional deep learning methods, it only requires less data to train a better model.

[0039] 2) Less work is required for labeling; compared to traditional deep learning methods, fewer labeled data are needed to train a better model.

[0040] 3) Semi-supervised learning did not increase the complexity of the model. Although the offline training time was longer, the online running speed remained relatively fast.

[0041] 4) It has strong versatility and can transfer the concept of semi-supervised learning through synthetic data to other medical image data analysis. Attached Figure Description

[0042] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0043] Figure 2 This is a diagram illustrating a method for training a synthetic model provided in an embodiment of the present invention;

[0044] Figure 3 This is the discriminator structure of the GAN used in the embodiments of the present invention;

[0045] Figure 4 The generator structure of the GAN used in the embodiments of the present invention;

[0046] Figure 5 This is the severity category control structure in the embodiments of the present invention;

[0047] Figure 6 This is a diagram illustrating a method for training a supervised target detection model provided in an embodiment of the present invention.

[0048] Figure 7 This is a diagram illustrating a method for training a semi-supervised object detection model provided in an embodiment of the present invention.

[0049] Figure 8 This is a schematic diagram of the periodontitis diagnosis results provided in an embodiment of the present invention. Detailed Implementation

[0050] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0051] Please see Figure 1 This invention provides a semi-supervised panoramic X-ray target detection method based on synthetic data, applied to the auxiliary diagnosis of periodontitis. Addressing the problem of insufficient image data and inadequate data annotation in existing methods leading to poor accuracy of trained diagnostic models, this invention aims to achieve better diagnostic results. The specific implementation includes the following steps:

[0052] Step 1: Obtain the panoramic X-ray dataset, train the synthetic model, and control the synthetic model to generate class-balanced samples; the method for training the synthetic model is as follows: Figure 2 As shown. It includes the following sub-steps:

[0053] Step 1.1: Prepare a small dataset of labeled panoramic X-ray images as real images, labeled as periodontitis. Encode each target's disease label category in each sample separately, and then weight the results to obtain the overall disease label, which serves as the category label for the entire sample. Figure 2 In this model, category control and category prediction are involved. Category control refers to generating samples of the corresponding class given a Severity Coefficient (SC) value; category prediction refers to predicting the class (SC value) of the generated samples. During training, both control and prediction need to be performed simultaneously to facilitate adversarial training. After the model is trained, only the generator model needs to be used to generate the category control function. This class is defined as the Severity Coefficient (SC).

[0054]

[0055] In the formula C i W represents the number of categories corresponding to the i-th target, and N is the number of targets in a single sample; i A severity coefficient is a weighted coefficient based on the number of categories. Since a panoramic radiograph may contain multiple dental targets, each with a corresponding category label, it is difficult to define the overall label. The severity coefficient is defined as a weighted average of individual values ​​to represent the overall value, converting the category label into a continuous quantity. In this research question, the categories of dental targets are healthy, mild, moderate, and severe. The definition ensures that a higher severity coefficient value indicates a more severe overall condition; that is, from mild to severe, "healthy" has a weight close to 0, and "severe" has a weight close to 1.

[0056] Step 1.2: Train a deep generative adversarial network (GAN) using the dataset. The goal is to synthesize high-quality results with high realism while satisfying the classification requirements.

[0057] The loss function of a typical GAN ​​is defined as follows:

[0058]

[0059] Among them, E z～p(z) [] represents expectation, P z (z) is the distribution of the values ​​of z, P data(x) represents the distribution of x as real samples (data); z is the input random vector, and x is a real sample; the discriminator D is trained to maximize the distribution distance between real samples and generated samples, while the generator G is trained to minimize the distribution distance to mislead the discriminator. By training the objectives of the two networks sequentially, i.e., adversarial training, the generator is eventually trained into a generative model that meets the requirements of high quality and controllability. Theoretically, the generator and discriminator of a GAN can be deep neural networks with arbitrary connection structures and dimensions, but different structures will affect the quality of the synthesized results.

[0060] The definition of the GAN model used includes the structure of the generator and discriminator, as well as the definition of the training loss function; improvements have been made in each part to generate high-quality, controllable results.

[0061] Conventional GANs take a random vector Z as input, sampling a Z as input during both training and generation. However, in the proposed GAN, Z is sampled and mapped to S during training; during generation, no mapping is needed, and a random vector S is directly sampled as input. The advantage of this is that the mapped S has better normality, resulting in more stable generation quality compared to the original method's direct Z input.

[0062] Discriminator structure as follows Figure 3 As shown in the diagram, compared to traditional network structures, the proposed discriminator has two main branches. The input sample to be discriminated is processed through the upper branch, passing through several residual downsampling blocks (D blocks) to reduce the feature map size; and through the lower branch, passing through a simple nearest neighbor downsampling block (P block) to further reduce the feature map size. After each downsampling stage, the two branches perform feature map addition, fusing the results as the input for the next stage. The final fused features are then processed through another residual downsampling block before the prediction result is output. In the diagram, FM refers to the feature map. The P structure uses a 2x2 nearest neighbor downsampling operation. In the D structure, the input first undergoes a 1x1 convolution, then splits into two paths: one branch performs two 3x3 convolutions, and the other performs a 1x1 convolution. The two paths then perform feature map addition, and finally, a 2x2 nearest neighbor downsampling operation is performed before the output. The biggest innovation is the proposed cross-layer connection structure, shown in the diagram as the downsampling channel (green P module) at the bottom of the network structure. Compared to traditional structures, this structure can better adapt to the adversarial learning needs of different training levels under high-resolution tasks.

[0063] Generator structure as follows Figure 4As shown in the diagram, compared to traditional network structures, the proposed generator mainly consists of two parts: an upper mapping module that maps the input vector Z to a style vector S through multiple fully connected layers, and a lower main module that controls the input generation result using the style vector. In the main module, the low-resolution input is expanded through multiple upsampling feature extraction blocks (B blocks), a process controlled by the components of the style vector S. Then, the output of each stage of the B block is output as OUT1, ..., OUT7 through the style mapping output block (R block), and the sum of all outputs is the final result. In the B block structure, the input feature map is upsampled, then undergoes two 3x3 convolutions, and finally outputs. In the R block structure, the input feature map undergoes a 3x3 convolution, then upsamples, and finally outputs. Since the input to output requires seven upsampling operations (depending on the specific engineering problem), each upsampling stage needs a corresponding style vector for control. S1-S7 are the seven components of the style vector S. The biggest innovation is the proposed fully connected mapping structure, shown in the diagram as the mapping channel from the random vector at the top of the network structure to the style vector. Compared to traditional structures, this structure can separate the input random vector from the upsampling structure and map it to style control vectors of different levels through multiple fully connected layers.

[0064] Since semi-supervised learning requires severity categories for synthetic samples, severity control is introduced into the GAN structure. Its mechanism is as follows: Figure 5 Based on the generator and discriminator structures, and given a model training loss function that includes SC categories, provide... Figure 5 The model training process is as follows:

[0065] Given the SC value, along with the style vector S and a constant input (which can be arbitrary, such as a matrix of all 1s, as long as the training and prediction inputs are consistent), these inputs are used as generator inputs. The generator then processes these inputs to output synthetic samples (fake samples with SC categories). These synthetic samples are input to a discriminator, which determines their authenticity and SC category. The discriminator's values ​​are then compared with the given SC value to calculate a loss function, which guides the training of both the generator and discriminator. Finally, a generator capable of controlling the synthesis of a given SC value is obtained.

[0066] In the implementation, the severity is controlled using a loss function and regularization, as follows:

[0067] The original definition of the loss function in GAN is incorporated into the severity coefficient control term, as shown below for the losses of the generator and discriminator, respectively:

[0068]

[0069]

[0070] In addition, a path length regularization term is introduced for training. This term is used to interpolate different input vectors. If the result satisfies a linear distance relationship, it can be inferred that the quality of the synthesized image is high. The definition is as follows:

[0071] Reg spl =λ(||G(z')-y||2-α) 2

[0072] In the formula, z' is the sampling of the input style vector; y is the image generated from z input; α is the average path constant of the sample set; and λ is the coefficient of the set regularization intensity.

[0073] In summary, the loss function of this GAN model is defined as follows, which can meet the requirements of high-quality, controllable synthetic data:

[0074] L = Loss D +Loss G +Reg spl

[0075] Step 1.3: Apply the trained deep generative adversarial network to synthesize synthetic data that is more than 10 times larger than the original labeled dataset, while also ensuring class balance in the synthetic data.

[0076] Step 2: Obtain the panoramic X-ray dataset and diagnostic labels, and train a supervised model for object detection and diagnosis; the method for training the supervised object detection model is as follows: Figure 6 As shown. It includes the following sub-steps:

[0077] Step 2.1: Prepare a small dataset of labeled panoramic X-ray images. The labels include the location information of each tooth and the category of periodontitis severity. Encode each label according to the sample to obtain the target tensor that the network model needs to fit.

[0078] Step 2.2: Based on the deviation between the fitted target and the network output target, train the network parameters to make the deep network fit the target. The deep target detection model shown in the figure is a one-stage detection model, which directly outputs the encoding of the predicted target through an end-to-end fitting. The network connection structure can be a convolutional neural network.

[0079] Step 3: Obtain weak labels for the synthetic samples, and train the diagnostic model using a semi-supervised object detection method on the entire unlabeled dataset to achieve more accurate diagnostic performance; the method for training the semi-supervised object detection model is as follows: Figure 4 As shown. It includes the following sub-steps:

[0080] Step 3.1: Using the target detection and diagnostic model obtained in Step 2, apply it to the unlabeled samples synthesized in Step 1 to obtain weak labels for the samples.

[0081] Step 3.2: Train the semi-supervised object detection model. Introduce noise into the input samples and use consistency regularization to constrain the model to convergence on a large dataset with weak labels, resulting in a preliminary semi-supervised detection model. The noise introduced here is achieved by randomly shifting, flipping, Gaussian blurring, etc., of the input samples to augment the data's shape without altering its semantics. The consistency regularization constrains the model's network parameters by reducing the distance between the predicted values ​​of noisy samples and those of noise-free samples. Preferably, the network structure of the semi-supervised object detection model is twice the number of network channels of the object detection and diagnosis model in Step 2.

[0082] Step 3.3: Using the semi-supervised detection model obtained in Step 3.2 as the object detection and diagnosis model in Step 3.1, weak labels are reacquired, and Steps 3.1 and 3.2 are repeated. This iterative process is repeated several times until the model fully converges, resulting in the final semi-supervised detection model. Here, the number of iterations is 5, meaning that the model reaches convergence after reacquiring weak labels 4 times after replacing the model.

[0083] Step 4: Deploy the trained target detection and diagnostic model on the software backend of the terminal device, and import the patient's panoramic X-ray data to make a diagnosis. The final online demonstration of the diagnostic model is shown in the image below. Figure 8 As shown.

[0084] Although this embodiment is applied to the detection and diagnosis of periodontitis, it does not mean that the invention is only applicable to the diagnosis of periodontitis; the invention is also applicable to other auxiliary medical fields. Simply replace the training dataset with a dataset from the corresponding field and retrain the network to obtain the desired results.

[0085] The above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A semi-supervised panoramic X-ray object detection method based on synthetic data, characterized in that, Comprising the following steps: Step 1, obtaining a panoramic X-ray data set, training a deep synthesis model, and controlling the deep synthesis model to generate a class-balanced sample; Comprising the following sub-steps: Step 1.1, preparing a small amount of labeled panoramic X-ray data set, encoding each target disease label category in each sample respectively, and obtaining the overall disease label as the class label of the entire sample after weighting; Step 1.2, training the deep synthesis model with the data set, requiring the synthesis result to meet the class while synthesizing high-quality results with high authenticity; The deep synthesis model is an improved generative adversarial network, including a generator and a discriminator, and the specific processing process of the generative adversarial network is: given a class label, i.e., a severity coefficient SC value, together with a style vector S and a constant input, as inputs to the generator, these input quantities are output through the operation of the generator to synthesize a sample, i.e., a false sample with an SC category, wherein the style vector S is obtained by mapping a random vector Z, and the constant input is arbitrary; the synthesized sample is input into the discriminator to obtain the discrimination values of its authenticity and SC category respectively, the discrimination values are calculated with the given SC value to obtain a loss function, which guides the training of the generator and the discriminator, and finally obtains a generator that can control the synthesis of a given SC value; The generator includes two parts, the upper side is a mapping module that maps the input vector Z to the style vector S through multiple layers of full connection, and the lower side is a main module that controls the input of the style vector to generate the result, in the main module, the input of low resolution size is expanded in resolution through multiple layers of B block, and the process is controlled by the components of the style vector S respectively, then the output of each stage B block is output as OUT1, … OUTn through the R block, and the sum of all is the output result, n represents the number of components of the style vector; wherein in the structure of the B block, the input feature map is up-sampled, then two 3X3 convolutions are performed, and finally output; in the structure of the R block, the input feature map is 3X3 convolved, then up-sampled, and finally output; The discriminator has two branches, the input to be discriminated sample passes through the upper branch to reduce the size of the feature map through a plurality of residual down-sampling blocks D blocks; through the lower branch, the size of the feature map is reduced through a simple nearest down-sampling block P block; the two branches are fused after each down-sampling stage through feature map addition, and the fused feature is output as a prediction result after passing through a residual down-sampling block; wherein the P block adopts a 2X2 nearest down-sampling operation, and in the structure of the D block, the input is first 1X1 convolved, then divided into two paths, one path is twice 3X3 convolved, the other path is 1X1 convolved, then the two paths are added to the feature map, and finally a 2X2 nearest down-sampling operation is performed to output; Step 1.3, applying the trained deep synthesis model to synthesize a synthetic data set K times larger than the original labeled data set, while satisfying the class balance of the synthetic data; Step 2, obtaining a panoramic X-ray data set and diagnostic annotation, training a deep target detection network model; Step 3, obtain weak labels of the synthetic samples, and train a diagnosis model using a semi-supervised object detection method on the entire unlabeled dataset to achieve more accurate diagnosis performance; The following sub-steps are included: Step 3.1, using the object detection diagnosis model obtained in step 2, implement it on the unlabeled samples synthesized in step 1 to obtain weak labels; Step 3.2, train the semi-supervised object detection model, introduce noise to the input samples, and constrain through a consistency regularization term, so that the semi-supervised object detection model is trained on the dataset with weak labels until convergence, obtaining a preliminary semi-supervised detection model; Step 3.3, use the semi-supervised detection model obtained in step 3.2 as the object detection diagnosis model in step 3.1 to re-obtain weak labels, and repeat steps 3.1 and 3.2, this iteration process is repeated several times until the model converges completely, obtaining the final semi-supervised detection model; Step 4, deploy the trained object detection diagnosis model in the software backend of the terminal device, and import the panoramic X-ray data of the patient to make a diagnosis.

2. The semi-supervised panoramic X-ray object detection method based on synthetic data of claim 1, wherein: The specific implementation of step 1.1 is as follows: The category label is defined as a severity coefficient SC: wherein is the number of classes corresponding to the th target, N is the number of targets in a single sample; is the number of classes corresponding to the th target, N is the number of targets in a single sample; 3. The semi-supervised panoramic X-ray object detection method based on synthetic data of claim 1, wherein: The specific calculation process of the loss function is as follows: The original definition of the loss function of GAN is integrated into the severity coefficient control term, as follows for the loss of the generator and discriminator respectively: where E z~p(z) [ ] is the expectation, P z (z) is the distribution of z values, P data (x) is the distribution of x values for real data; is the input random vector, is the real data; the discriminator D training objective is to maximize the distribution distance between real and generated samples, while the generator G training objective is to minimize the distribution distance to mislead the discriminator; A path length regularization term is additionally introduced for training, which is to interpolate between different input vectors to obtain a result that satisfies the linear distance relationship, which is defined as follows: wherein is a sampling of the input style vector; is input generated imagery; is an average path constant for the sample set, is a coefficient for the set regularization strength; In summary, the loss function of the improved generative adversarial network is defined as follows: 。 4. The semi-supervised panoramic x-ray object detection method based on synthetic data of claim 1, wherein: The specific implementation of step 2 includes the following sub-steps: Step 2.1, prepare a small amount of labeled panoramic X-ray dataset, encode each label according to the sample to obtain the target that the deep object detection network model needs to fit; Step 2.2, according to the deviation between the fitting target and the output target of the deep object detection network model, train the network parameters to make the deep object detection network model fit.

5. The semi-supervised panoramic X-ray object detection method based on synthetic data of claim 1, wherein: The deep object detection network model is a one-stage detection model or a two-stage detection model based on a convolutional neural network.

6. The semi-supervised panoramic x-ray object detection method based on synthetic data of claim 1, wherein: The network model structure of the semi-supervised object detection model is the same as the deep object detection network model in step 2.

Images